Deep learning-based assessment of facial structure changes after oral rehabilitation

Accurate prediction of postrestoration facial changes in edentulous patients remains challenging due to intricate soft-tissue dynamics, limiting preoperative planning precision. This study presented FacePointNet, a bidirectional deep learning framework for simulating localized facial alterations after dental rehabilitation. Sixteen edentulous patients underwent preoperative/postoperative three-dimensional facial imaging. The model integrates dual-registration subnetworks to analyze bidirectional geometric transformations between baseline and postoperative states, optimized via composite loss functions (geometric consistency and topology preservation). Validation employed fourfold cross-validation, with outcomes assessed using Chamfer distance (CD), Euclidean landmark errors, and blinded Visual Similarity Scoring by three clinicians. FacePointNet achieved submillimeter accuracy in global morphology (CD, 0.87 mm) and localized errors of 3.80 to 5.98 mm (nasal tip, 2.17 mm; oral commissures, ≤6.88 mm). Predicted versus actual outcomes showed no statistical divergence (p>0.05), with expert evaluators confirming high visual fidelity, particularly in nasolabial fold reproduction. This study established the first artificial intelligence-driven system for patient-specific prosthodontic outcome projection, demonstrating clinical-grade precision in critical aesthetic zones. FacePointNet advances personalized dental rehabilitation by enabling evidence-based preoperative counselling and data-driven treatment design, addressing a critical gap in aesthetic prosthodontics.